Last updated on April 1st, 2026 at 11:28 pm

image from @birdabo on Twitter/X

Print designers have always had reactive text (reactive until set). Words flow around images, wrap into columns, and fit the shape of the page, because the layout engine understands the relationship between text and space. The web was supposed to do this too. For thirty years, it mostly couldn’t.

A new open-source library called Pretext has apparently closed that gap.

Lou, the engineer behind React Motion, the co-creator of ReasonML, and a key contributor to Facebook Messenger’s frontend, released Pretext this week, a pure JavaScript/TypeScript library that measures and lays out multiline text without touching the DOM. The result is text that flows around images, wraps into columns, and fits irregular shapes the way magazine and newspaper layouts have done for a century. In a browser. At 120 frames per second.

What Pretext Actually Does

The core insight is deceptively simple. Every time a browser needs to figure out where text goes on screen, it triggers what’s called a layout reflow, one of the most computationally expensive operations in web rendering. Resize a window, change a font size, add a word, and the browser recalculates the position of every element on the page. This is why complex web layouts stutter, why text-heavy interfaces feel sluggish, and why no one has successfully replicated print-quality typography on the web at scale.

Pretext implements its own text measurement logic using the browser’s font engine as ground truth, then performs all layout calculations as pure arithmetic. No DOM queries. No reflow. The library does a one-time measurement pass, segments the text, applies line-breaking rules with full internationalization support, and returns layout data that can be rendered to DOM, Canvas, SVG, or (soon) server-side.

The technical claim is 500x faster than DOM-based measurement. The demos, which include text flowing dynamically around animated objects, tight-fitted message bubbles, and full editorial spreads with obstacle-aware routing, demonstrate the tech isn’t just hype.

Why This Matters Beyond the Demo Reel

It frees designers. Digital media has spent two decades trying to make web content look as good as print, and mostly failing. Pretext doesn’t solve that problem completely. It doesn’t handle images, color, or overall page composition, but it removes the single biggest blocker: the inability to do fast, flexible, high-fidelity text layout in the browser.

But the more interesting story is what Pretext represents as a pattern.

Cheng Lou built it at Midjourney, an AI art generator. The library’s own documentation describes its iteration method as “AI-friendly.” Lou trained models against text measurement data to refine the library’s accuracy. This is a machine learning engineer applying ML methodology to a systems-level web problem that CSS working groups have been nibbling at for decades.

This is the pattern enterprise technology leaders should be watching. The most consequential AI-adjacent work isn’t happening in chatbots or image generators. It’s happening when engineers who’ve been steeped in ML methodology start looking at infrastructure problems that the rest of the industry has accepted as permanent limitations.

What It Doesn’t Do (Yet)

Pretext is explicit about its current scope. It handles text measurement and layout, not rendering, not styling, not accessibility. It’s designed to be a foundation that other libraries build on top of. It doesn’t handle variable fonts across different weights in a single run. It doesn’t do vertical writing modes yet. It targets the common text setup and acknowledges its boundaries clearly.

The library is also brand new. It has 5,600 GitHub stars in its first days, which signals serious developer interest, but production adoption at scale is a different conversation. Whether Pretext becomes foundational infrastructure or remains an impressive proof of concept depends on whether the ecosystem builds on top of it.

The Enterprise Signal

For B2B technology leaders, the lesson of Pretext itself may not seem immediately relevant. Most enterprise applications aren’t pushing the boundaries of text layout. But the story underneath it is.

The AI talent pipeline is starting to produce engineers whose default problem-solving toolkit includes model training, iterative optimization against ground truth, and the assumption that systems-level performance problems can be rethought from first principles rather than patched incrementally. These engineers are leaving pure-play AI companies and applying those instincts to the unsexy plumbing of the technology stack.

The companies that will benefit most are the ones hiring engineers who think like Cheng Lou, people who look at a thirty-year-old problem and ask whether it’s actually a limitation, or just a habit, apply ML or AI and a little imagination, and reinvent something that has existed as is for a long time.

Pretext is open source and available on GitHub at github.com/chenglou/pretext. Live demos are at chenglou.me/pretext.